This research investigates how prompts can be extracted from customized LLMs, threatening intellectual property and business models based on prompt engineering.

Key findings:

• Identifies mechanisms behind prompt extraction vulnerabilities in large language models
• Demonstrates how attackers can steal valuable prompt designs that represent significant IP
• Analyzes the risk to businesses relying on LLM customization (like OpenAI's GPTs)
• Proposes defense strategies to protect prompt-based services

Business Impact: Companies investing in prompt engineering as a business strategy need robust security measures to protect their intellectual property from extraction attacks.

Why Are My Prompts Leaked? Unraveling Prompt Extraction Threats in Customized Large Language Models